Holographic three-dimensional bioassay method based on depth label coding and attention mechanism

Abstract

The invention discloses a holographic three-dimensional bioassay method based on depth label coding and an attention mechanism. The invention relates to a holographic biometric platform employing depth label coding and end-to-end attention-enhanced three-dimensional reconstruction. The platform can enhance the depth of field and volume detection capability, and significantly improve the detection sensitivity and accuracy. A depth label coding strategy is integrated into a neural network architecture based on an attention mechanism, and three-dimensional volume reconstruction is converted into a dual-task framework combining bounding box detection and depth classification. And the limitation of a traditional generation type holographic reconstruction algorithm is overcome, so that accurate real-time three-dimensional positioning is realized. The holographic three-dimensional space positioning is beneficial to expanding field depth application and improving the density of the countable microspheres, chloramphenicol can be detected in a wide dynamic range, and compared with the traditional bioassay based on a two-dimensional hologram, the sensitivity is improved.

Description

Technical Field

[0001] This invention relates to the interdisciplinary field of immunoassay detection technology, optical engineering, and computer image analysis technology, and in particular to a holographic three-dimensional bioassay method based on depth tag encoding and attention mechanism. Background Technology

[0002] With the escalating concerns about food safety, environmental pollution, and public health, the demand for highly sensitive and rapid detection technologies for low-concentration target analytes, especially trace antibiotics, small molecule drugs, and pathogens, is growing. Immunoassay, an analytical technique based on antigen-antibody specific recognition reactions, has been widely applied in food safety testing, environmental monitoring, and clinical diagnosis due to its high selectivity and sensitivity. Traditional immunoassay methods, such as enzyme-linked immunosorbent assay (ELISA) and chemiluminescent immunoassay (CLIA), while possessing good detection performance, still face numerous challenges. For example, these methods typically require complex operating procedures, expensive equipment, long reaction cycles, and specialized operators, limiting their widespread adoption in rapid on-site testing or portable devices. In recent years, microsphere probe-based immunosensing methods have attracted widespread attention due to their ease of operation, rapid response, and integrability. Polystyrene microspheres (PSM) or magnetic nanoparticles (MNP) are often used as carriers, with antibodies or antigens linked through surface chemical modification to achieve target analyte recognition and capture. However, how to achieve rapid, efficient, and high-throughput quantitative identification of these microsphere probes remains a bottleneck in the development of current immunosensing technologies.

[0003] In signal readout and analysis, bioassays based on microscopic imaging technology have attracted much attention due to their high sensitivity and accuracy in biosensing applications. However, traditional bright-field microscopy is inherently limited by its narrow field of view and shallow depth of field, which restricts the analysis of large-volume samples and 3D reconstruction. In contrast, lensless holographic imaging has emerged as a highly promising imaging technique, capturing the interference pattern between a reference wave and an object wave using a complementary metal-oxide-semiconductor (CMOS) sensor. Its main advantages include no need for focusing, non-invasiveness, extended depth of field, wide field of view, and 3D reconstruction capabilities, making it an ideal technique for high-sensitivity imaging. Applying the backpropagation algorithm to the captured 2D hologram can accurately reconstruct the 3D depth distribution of suspended particles in solution, thus overcoming the inherent limitations of traditional microscopy. This method is of great significance for applications requiring high-sensitivity detection, spatial localization, dynamic particle tracking, and quantitative analysis of actual samples.

[0004] Currently, there are two main techniques for reconstructing 3D structures from 2D holograms: optical backpropagation algorithms based on physical principles and data-driven deep learning algorithms. Classical optical reconstruction methods include Rayleigh-Sommerfeld backpropagation, angular spectral propagation, and Fresnel transform techniques, all based on well-established wave optics principles. These methods are computationally efficient and do not require large-scale training datasets. However, their strong dependence on accurate initial estimates and prior knowledge limits their applicability in dynamic imaging scenarios. Furthermore, their sensitivity to noise and poor adaptability to complex environments reduce their robustness in practical applications. In contrast, deep learning-based 3D holographic reconstruction offers greater flexibility. By computationally compensating for parameter mismatches, it reduces reliance on precise optical configurations and minimizes sample preprocessing. These models typically require minimal post-processing and achieve higher reconstruction accuracy. However, their performance heavily depends on the quantity and quality of training data, often requiring accurate datasets with depth annotations, which are difficult to obtain. Additionally, heterogeneous media and complex environmental factors can affect the model's generalization ability and prediction accuracy. Despite some progress in recent years, the practical application of deep learning to large-volume holographic particle imaging remains under development. Models trained on simulated data often struggle to generalize to real-world scenarios involving complex refractive environments. Therefore, there is an urgent need to develop a computationally efficient and accurate reconstruction method for applications in fields such as food safety testing, environmental monitoring, and biomedical diagnostics. Summary of the Invention

[0005] This invention provides a holographic three-dimensional biometric method based on depth tag encoding and attention mechanism.

[0006] The technical solution of this invention is as follows: A holographic three-dimensional biometric method based on depth label encoding and attention mechanism, the method comprising the following steps: Step 1: First, the target analyte is co-incubated with trans-cyclooctene-antibody conjugate (TCO-Ab) and fully antigen-modified magnetic nanoparticles (MNP-BSA-CAP) to perform an immunoassay. After the reaction, magnetic separation is performed, and the precipitate, which is TCO-Ab bound to MNP, is collected. Subsequently, this precipitate is subjected to a bioorthogonal click cycloaddition reaction with tetrazine-functionalized polystyrene microspheres (PSM-Tz) at room temperature. After the reaction, magnetic separation is performed again, and the supernatant obtained from this separation is finally collected and added to a custom-designed deep-grooved slide for holographic imaging analysis using a portable lensless holographic imaging microscope. The concentration of polystyrene microspheres in the supernatant is linearly correlated with the original concentration of the target analyte.

[0007] Step 2: Using a random particle generation algorithm, a particle distribution with precise three-dimensional coordinate labels was created; then, the angular spectrum method combined with double Fourier transform was used to simulate the forward digital holographic propagation process, and Gaussian noise of different intensities was added to simulate real-world holographic data, resulting in a dataset containing three-dimensional coordinates and corresponding two-dimensional holograms. Step 3: An end-to-end network architecture based on the attention mechanism (3D_EEA network) is used to complete the 3D reconstruction of the 2D holographic image. The coordinate localization and depth regression of particles at different depths are performed by using a 3D depth dataset and combining a deep learning model to complete the localization and reconstruction of particles in 3D space. Step 4: Utilizing the linear relationship between the number of polystyrene microspheres and the concentration of the target analyte, the polystyrene microspheres are located and quantitatively analyzed through holographic three-dimensional reconstruction. Finally, the number of polystyrene microspheres is calculated to obtain the concentration information of the target analyte.

[0008] Preferably, step 2 specifically includes the following: 2.1 A synthetic particle field is generated in three-dimensional space using a random particle generation algorithm; the spatial coordinates (X, Y, Z) of the particle center are generated according to the Poisson distribution, while allowing up to 30% particle overlap along the Z-axis to simulate the random and independent distribution of particles in a large volume space under real conditions. The coordinate information of each particle includes the bounding box coordinates (x, y, w, h) and the normalized depth label z ∈ [0,1], where z is discrete into K=100 categories, which constitute the depth classification label; 2.2 The angular spectrum iterative method is used to backpropagate the 3D particle distribution to the XY plane and synthesize the corresponding 2D hologram. Specifically, this includes: initializing the light field, adding random particles, performing Fourier transform and phase modulation using the angular spectrum transfer function, and recording the superimposed holograms. To simulate a real imaging environment, Gaussian noise with different signal-to-noise ratios is added to the synthesized holograms. Finally, a dataset containing 3D coordinate labels and corresponding 2D holograms is constructed for training an end-to-end 3D reconstruction model. In step 2, the dataset for the two-dimensional holograms includes XML files with particle localization and depth label information, as well as corresponding digital holograms. A total of 25,000 samples were generated and divided into training, validation, and test sets in a 7:2:1 ratio. To simulate real-world scenarios and augment data, Gaussian noise with different signal-to-noise ratios (SNR) was added to 5,000 images in the training set. Additionally, 200 experimentally acquired holograms (512×512 pixels) were included in the test set to evaluate the predictive performance of the 3D_EEA model on both simulated and real data.

[0009] More preferably, in step 2.2, the angular spectrum iterative algorithm propagates the three-dimensional particle distribution into the two-dimensional holographic plane. The light source wavelength (λ) is set to 633 nm, the imaging sensor pixel size (Δx, Δy) is set to 1.85 μm, the refractive index (n) of the propagation medium is 1.33, the propagation distance (z) is 0.5 mm-1.5 mm, the calculation grid is set to 512×512 pixels, and the angular spectrum transfer function is defined as: H(fx, fy) = exp( j*2π * (z / n) * sqrt( (1 / λ)² - fx² - fy² )); the simulated particle diameter is the same as the actual particle diameter, which is 6 μm.

[0010] More preferably, in step 2.2, the Gaussian noise with different signal-to-noise ratios is added, specifically 10 dB, 15 dB, and 20 dB.

[0011] Preferably, in step S3, the 3D_EEA network adopts an encoder-decoder architecture and embeds an attention mechanism. Its specific structure and data flow are as follows: The feature extraction encoder (LR module) consists of four lightweight residual blocks connected in series and serves as the backbone of the network. It is responsible for extracting multi-level features from the input two-dimensional hologram (size: 512×512) and outputting a high-dimensional feature tensor. The multi-scale feature fusion module consists of an upsampling path UF module and a downsampling path DF module. This module receives multi-level outputs from the encoder and forms a feature pyramid structure through cross-layer cascading. Finally, it outputs an enhanced feature (size 128×128, number of channels 256) that integrates multi-scale information. This feature is used to fuse the deep semantic features and shallow spatial detail features output by the encoder to improve the perception of particles at different scales. The Convolutional Attention Module (CBAM) adaptively redefines the channel and spatial dimensions of the fused feature map F2 through the channel attention submodule and the spatial attention submodule, outputting the redefinition features, enhancing the network's attention to particle regions, and suppressing the interference of complex background noise. The output head consists of a deep regression branch and a counting branch. Based on the redefinition features, the network finally outputs the three-dimensional coordinates (x, y, z), confidence score, and total number of particles in the image for each particle. The network training adopts a supervised learning approach, using the dataset constructed in step 2. The loss function includes target loss, bounding box regression loss, and depth regression loss. This network architecture design enables holographic three-dimensional spatial localization and quantitative analysis of particles.

[0012] More preferably, in step S3, the channel attention submodule first generates two one-dimensional (1D) channel descriptors through two operations: global average pooling and global max pooling. These descriptors are then processed by a shared multi-layer perceptron (MLP) to generate channel attention weights. This MLP contains two convolutional layers with a kernel size of 1. The specific process is as follows: the first layer reduces the channel dimension by a reduction factor r=16, followed by ReLU activation; the second layer restores the original channel dimension; and the final output weights are normalized using the sigmoid function.

[0013] More preferably, in step S3, the spatial attention submodule takes the channel-optimized features as input and first performs average pooling and max pooling operations along the channel dimension to generate two two-dimensional (2D) spatial feature maps. These feature maps are then concatenated and processed by a standard convolutional layer (7×7 kernel size) to generate a spatial attention map. The specific parameters of this convolutional layer are as follows: input channels = 2; output channels = 1; kernel size = 7×7; stride = 1; padding = 3; the resulting spatial weights are also normalized using the sigmoid function.

[0014] Preferably, in step S3, the network training parameters of the end-to-end 3D reconstruction model include: 200 training iterations, an initial learning rate of 0.01, a batch size of 16, using the Adam optimizer, a momentum of 0.937, and a weight decay of 0.0005.

[0015] More preferably, in step S3, the loss function is specifically: ; in, For bounding box regression loss, For deep classification, cross-entropy loss, The target confidence binary cross-entropy loss is used. The value is 0.25. The value is 1.0. The value is set to 1.0, and the weight parameter is determined through empirical parameter tuning.

[0016] Preferably, the portable lensless holographic imaging microscope in step S1 includes a coherent light source, an optical slit, a CMOS image sensor, and a 3D-printed dark-field shell, wherein the wavelength of the coherent light source is 532 nm, the pixel size of the CMOS image sensor is 1.85 μm, and the resolution is 4032×3036; the depth of the recessed slide is 0.8 mm, the diameter is 15 mm, and the thickness is 1 mm. PSM consists of carboxylated polystyrene microspheres with a diameter of 6 μm, and the surface is modified with Tz click molecules for click reactions. MNPs are carboxylated magnetic nanoparticles with a diameter of 150 nm, and their surfaces are modified with the complete antigen of the target analyte. The target analytes are small molecule antibiotics, including chloramphenicol (CAP), neomycin, clarithromycin, and other antibiotics, as well as pathogenic bacteria. The holographic immunoassay method has a linear detection range of 5 pg / mL to 100 ng / mL for CAP, with a coefficient of determination R² of 0.988.

[0017] Preferably, the time required for the holographic immunoassay method in step S1 is 40-50 minutes, and the reaction conditions for the competitive immunoassay are: 20-40 ℃, with a reaction time of at least 10 min.

[0018] Compared with the prior art, the beneficial effects of the present invention are: This invention addresses the fundamental technical bottlenecks faced by traditional microscopic imaging techniques in complex biological detection. Firstly, addressing the problem of narrow depth of field in traditional microscopes, which necessitate mechanical scanning for observing large-volume samples, this solution captures full-depth information in a single pass using lensless holographic imaging. It innovatively introduces a physical model-based angular spectrum algorithm to construct a training dataset (step 2), which is then used to train an end-to-end deep learning network (step 3). This enables the system to directly extract the three-dimensional spatial distribution of particles from a single two-dimensional hologram, thus transforming the "large depth-of-field imaging advantage" of holography into "precise quantitative capability with large depth of field." Secondly, addressing the problem of inaccurate counting caused by impurities and overlapping diffraction rings in actual samples, the deep learning model in this solution effectively distinguishes between real signals and noise through attention mechanisms and three-dimensional spatial feature learning. Furthermore, it separates overlapping diffraction rings based on the physical constraints of light propagation, achieving accurate identification and counting of target microspheres in complex backgrounds. Steps 2 and 3 are not simply a combination of technologies, but rather a deep collaboration between physical models and data-driven algorithms that solves the core technical challenge of "reliably reconstructing three-dimensional quantitative information from mixed two-dimensional signals," ultimately ensuring the high sensitivity, high anti-interference ability, and high reliability of the detection method.

[0019] This invention restructures the 3D reconstruction task into a dual-task framework of bounding box detection and depth classification, replacing traditional schemes based on physical backpropagation or generative models, significantly improving computational efficiency and inference speed. A depth label encoding strategy is introduced, discretizing continuous depth values ​​into classification labels, overcoming the instability problem of depth estimation in regression methods.

[0020] Employing a lightweight residual structure and a multi-scale attention mechanism, this method enhances the ability to perceive weak particle signals in holograms while maintaining the number of model parameters, making it suitable for 3D localization in large depth-of-field and complex backgrounds. Holographic 3D spatial localization helps expand depth-of-field applications and increase countable microsphere density, enabling the detection of chloramphenicol over a wide dynamic range of 5 pg / mL to 100 ng / mL, with a 10-fold increase in sensitivity compared to traditional 2D hologram-based bioassays. It achieves end-to-end mapping from 2D holograms to 3D coordinates without iterative optimization or post-processing, making it suitable for real-time, high-throughput bioassay scenarios.

[0021] This invention utilizes lensless holographic imaging technology combined with an end-to-end attention mechanism 3D reconstruction algorithm for microsphere signal probe quantification, enabling rapid and sensitive detection of antibiotic residues. The designed detection system exhibits high sensitivity, strong specificity, ease of operation, and high detection efficiency, showing broad application prospects in food safety testing, environmental monitoring, and in vitro monitoring.

[0022] This invention achieves compatible detection of multiple sample types. Addressing the problem of existing methods detecting only one target, this invention optimizes the algorithm architecture and feature extraction methods, enabling the system to simultaneously meet the detection needs of multiple targets such as nucleic acids and proteins. Attached Figure Description

[0023] Figure 1 This is a schematic diagram of the lensless holographic imaging hardware and a schematic diagram of the holographic 3D reconstruction algorithm based on tag encoding and attention mechanism described in the embodiments of the present invention.

[0024] Figure 2 The diagram (a) shows the principle of the algorithm for constructing a digital holographic simulation dataset as described in this embodiment of the invention, and the flowchart (b) shows the principle of the holographic 3D reconstruction algorithm based on label encoding and attention mechanism, which includes a lightweight residual module, a convolutional attention mechanism module, an upsampling module, and a downsampling module.

[0025] Figure 3 The characterization metrics used to evaluate the holographic 3D reconstruction model as described in this embodiment of the invention include targetability loss, bounding box regression loss, depth regression loss, accuracy, and recovery rate.

[0026] Figure 4Compare the computational speed of the numerical propagation reconstruction algorithm and the 3D_EEA algorithm under the same hardware configuration and image size (512×512 pixels).

[0027] Figure 5 Different levels of Gaussian noise are added to the digitally simulated holograms described in this embodiment of the invention to simulate the effect of 3D reconstruction algorithms in a real environment.

[0028] Figure 6 This invention provides the coordinate localization and 3D reconstruction results of a holographic 3D reconstruction algorithm based on tag encoding and attention mechanism at different microsphere concentrations, as described in this embodiment.

[0029] Figure 7 This invention presents the 3D reconstruction results and top-view comparisons of the holographic 3D reconstruction algorithm based on tag encoding and attention mechanism used in measured holograms with different microsphere concentrations in this embodiment of the invention.

[0030] Figure 8 The holographic 3D reconstruction algorithm based on label encoding and attention mechanism described in this embodiment of the invention is used to detect the concentration response of chloramphenicol 2D holograms and compare the corresponding 3D reconstruction results.

[0031] Figure 9 In this embodiment of the invention, a holographic three-dimensional reconstruction algorithm based on tag encoding and attention mechanism is used to detect the concentration gradient response of chloramphenicol (a) and the standard curve (b).

[0032] Figure 10 Different antibiotics were used as interfering agents in the embodiments of the present invention to test the specificity of the antibiotic detection strategy of the present invention.

[0033] Figure 11 This invention uses freshwater fish as real samples to detect the effectiveness of antibiotics (a) and compares it with the ELISA detection method (b).

[0034] Figure 12 This is a schematic diagram of the convolutional attention mechanism module of the holographic 3D reconstruction algorithm based on label encoding and attention mechanism described in this embodiment of the invention. Detailed Implementation The present invention will be further described below with reference to the embodiments, but the scope of protection of the present invention is not limited to the scope described in the embodiments.

[0035] Example 1: Specific operating steps and key parameters of the 3D_EEA biosensor like Figures 1-11 As shown, this embodiment provides a holographic spatial localization biometric detection method based on depth-labeled encoding and end-to-end attention-enhanced 3D reconstruction, including: Step 1: First, the target analyte is co-incubated with trans-cyclooctene-antibody conjugate (TCO-Ab) and fully antigen-modified magnetic nanoparticles (MNP-BSA-CAP) to perform an immunoassay. After the reaction, magnetic separation is performed, and the precipitate, which is TCO-Ab bound to MNP, is collected. Subsequently, this precipitate is subjected to a bioorthogonal click cycloaddition reaction with tetrazine-functionalized polystyrene microspheres (PSM-Tz) at room temperature. After the reaction, magnetic separation is performed again, and the supernatant obtained from this separation is finally collected and added to a custom-designed deep-grooved slide for holographic imaging analysis using a portable lensless holographic imaging microscope. The concentration of polystyrene microspheres in the supernatant is linearly correlated with the original concentration of the target analyte.

[0036] Step 2: Jointly encode the particle position and depth information extracted from the 2D hologram: Use a random particle generation algorithm to generate a 3D particle distribution, including 2D spatial coordinates and depth normalized values. The 2D spatial coordinates generate bounding boxes for particle counting, while the depth values ​​provide regression labels for 3D reconstruction; thereby constructing a 3D reconstruction algorithm dataset.

[0037] Step 3: Develop an end-to-end network architecture (3D_EEA network) based on an attention mechanism to complete the 3D reconstruction of 2D holographic images. Using a 3D depth dataset and a deep learning model, coordinate localization and depth regression are performed on particles at different depths to complete the localization and reconstruction of particles in 3D space. This attention mechanism network focuses more on processing particle channel and spatial features to improve sensitivity to changes in particle depth. This network architecture design enables accurate 3D holographic reconstruction and quantitative analysis. Step 4: Utilizing the linear relationship between the number of polystyrene microspheres and the concentration of the target analyte, the polystyrene microspheres are located and quantitatively analyzed through holographic three-dimensional reconstruction. Finally, the number of polystyrene microspheres is calculated to obtain the concentration information of the target analyte.

[0038] Optionally, the deep label encoding constructs a dataset containing three-dimensional coordinates and corresponding two-dimensional holograms using bounding box coordinates (x, y, w, h), normalized classification labels (z), and confidence scores (c), which is then used to train a deep learning model.

[0039] Optionally, the end-to-end attention mechanism-based 3D reconstruction architecture includes a lightweight residual network module for image feature extraction, an upsampling and downsampling fusion network for multi-scale feature fusion, an attention mechanism module for enhancing the attention response to relevant spatial and channel features, and finally outputs the 3D localization and quantity of particles through the output layer.

[0040] Optionally, the portable lensless holographic imaging microscope includes a coherent light source, an optical slit, a CMOS image sensor, and a 3D-printed dark-field shell, wherein the wavelength of the coherent light source is 532 nm, the pixel size of the CMOS image sensor is 1.85 μm, the resolution is 4032×3036, and the depth groove slide has a depth of 0.8 mm, a diameter of 15 mm, and a thickness of 1 mm.

[0041] Optionally, the training parameters of the end-to-end 3D reconstruction model include: 200 training iterations, an initial learning rate of 0.01, a batch size of 16, using the Adam optimizer, a momentum of 0.97, a weight decay of 0.0005, and training on a hardware device set built with an NVIDIA GeForce RTX 3090 Ti GPU and an Intel i9-12900K CPU.

[0042] Optionally, the PSM is carboxylated polystyrene microspheres with a diameter of 6 μm, and the surface is modified with Tz click molecules for click reactions.

[0043] Optionally, the MNP is a carboxylated magnetic nanoparticle with a diameter of 150 nm, and its surface is modified with the complete antigen of the target analyte.

[0044] Optionally, the target analyte is a small molecule antibiotic or a pathogenic bacterium, wherein the antibiotic includes at least one of CAP, neomycin, and clarithromycin, and the holographic immunoassay method has a linear detection range of 5 pg / mL to 100 ng / mL for CAP, with a linear regression coefficient R² of 0.988.

[0045] Optionally, the holographic immunoassay method requires 40-50 minutes, and the competitive immunoassay reaction conditions are 20-40 ℃ and a minimum reaction time of 10 min.

[0046] Optionally, the pre-prepared TCO-Ab, MNP-BSA-CAP, and PSM-Tz are all stored at 4°C after preparation.

[0047] Example 2: Construction and Performance Evaluation of 3D_EEA Biosensor Model (1) Establishment of the 3D_EEA model dataset This invention proposes a novel method for generating a synthetic particle field in three-dimensional space using a random particle generation algorithm. The particle diameter is set to 6 μm, the XY plane of the particle field is defined as 512×512 pixels, and the particle depth (Z-axis) ranges from 0.5 mm to 1.5 mm. The normalized depth z∈[0,1] is discretized into K=100 categories. The coordinate information of each particle (including (x,y,w,h)) and the normalized depth label are saved in VOC format and used as annotation labels for the 3D_EEA model. Subsequently, a digital holographic simulator based on the angular spectrum iteration method is used to backpropagate the randomly generated 3D particles onto the XY plane to synthesize the corresponding two-dimensional hologram. The two-dimensional hologram generation process based on the angular spectrum iteration method includes initializing the light field, sequentially adding particles, propagating the light field using the angular spectrum method, and recording the superimposed holograms. Specifically, angular spectrum propagation involves performing a Fourier transform on the initial field to convert it to the frequency domain, then using a transfer function for phase modulation to simulate free-space propagation, and finally reconstructing the propagated optical field through an inverse Fourier transform.

[0048] Finally, a 3D_EEA model dataset was successfully constructed. The dataset includes XML files with particle localization and depth label information, as well as corresponding two-dimensional digital holograms. A total of 25,000 samples were generated and divided into training, validation, and test sets in a 7:2:1 ratio. To simulate real-world environments, Gaussian noise with different signal-to-noise ratios was added to 5,000 images in the training set. Furthermore, 200 experimentally captured holograms (512×512 pixels) were included in the test set to evaluate the predictive performance of the 3D_EEA model on both simulated and real data.

[0049] (2) Algorithm construction of 3D_EEA model The 3D_EEA model consists of a digitally simulated holographic reconstruction dataset and an end-to-end attention-based neural network for 3D volumetric reconstruction. The attention-based reconstruction network includes a feature extraction module (LR module), a multi-scale feature fusion module (UF module and DF module), an attention mechanism module (CBAM module), and classification and regression output modules, such as... Figure 12 As shown, CBAM includes channel attention and spatial attention submodules, refining the intermediate features of the hologram along two orthogonal dimensions. By dynamically suppressing irrelevant background information and enhancing significant particle cues, CBAM significantly improves the model's robustness in handling noisy and low-contrast holographic data, thus contributing to accurate 3D position regression during volumetric holographic reconstruction. The simulated hologram and its corresponding 3D position labels are input into the model for supervised training.

[0050] The model was optimized using the Adam optimizer with a momentum of 0.937 and a weight decay of 0.0005. The initial learning rate was set to 0.01 and adjusted using a cosine annealing scheduler. Training was performed in batches of 16 for 200 epochs. The model was implemented using the PyTorch 1.12 framework in a Python 3.8 environment, with supporting libraries including torch, torchvision, numpy, and opencv-python. All experiments were conducted on a workstation equipped with an NVIDIA GeForce RTX 3090 Ti GPU, an Intel i9-12900K CPU, and 32 GB of RAM.

[0051] The training curves for the first 100 batches show that the model loss steadily converges. Figure 3 Specifically, the target loss used to evaluate particle presence, the bounding box regression loss used to evaluate particle localization in the XY plane, and the depth regression loss used for Z-axis prediction all converged on both the training and test sets after approximately 60 batches of training. Furthermore, none of the losses showed an increasing trend after convergence, indicating no overfitting. The model achieved its highest detection precision of 0.963 on the 74th batch and a recall of 0.919 on the 99th batch, indicating low false positive and false negative rates. Under the same hardware environment and image size, the numerical propagation-based reconstruction algorithm improved computation speed by 32 times compared to the 3D_EEA algorithm (image size: 512×512 pixels). Figure 4 These results demonstrate that the model is robust and reliable in particle detection and 3D localization, confirming its potential in predictive applications.

[0052] To improve the realism of the simulated holographic data during training, we randomly added Gaussian noise of different intensities to the simulated dataset and evaluated the model's predictive performance under different signal-to-noise ratios (SNR). Figure 5 The results show that the model maintains reliable prediction accuracy when SNR > 15 dB. However, when SNR = 10 dB, although the overall particle count remains accurate, excessive noise causes some particles to have predicted position deviations. These findings confirm the model's robustness and prediction reliability at moderate noise levels.

[0053] Figure 6This paper demonstrates the prediction and 3D reconstruction performance of the model under different particle concentrations in simulated holographic data. The results show that the 3D_EEA model achieves a 100% detection rate at all concentrations. Although a depth prediction error occurred at the highest particle concentration, the overall detection accuracy remains excellent. To evaluate the model's generalization ability to real-world data, we applied the trained model to real holographic images containing different particle concentrations. The model successfully detected the particles and predicted their depth information. Figure 7 We projected the corresponding 3D reconstruction results onto a 2D plane and compared them with the original hologram. The predicted particle positions showed a high degree of spatial consistency with the original image, and the particle count was very close to the true value. Example 3: Preparation and Performance Evaluation of the 3D_EEA Biosensor for Detecting Chloramphenicol This embodiment provides a holographic three-dimensional biometric method based on depth tag encoding and attention mechanisms. See [link to documentation]. Figure 1 This embodiment is used to detect the concentration of chloramphenicol in the sample to be tested.

[0054] First, let's introduce the sources of the main reagents used in this embodiment: Carboxyl-functionalized polystyrene microspheres (PSM-COOH, 6 μm) were purchased from Bangs Laboratories, Inc. (USA).

[0055] Carboxylated magnetic nanoparticles (MNP) 150 -COOH was purchased from Ocean Nano-Tech (USA).

[0056] N-hydroxysuccinimide ester-polyethylene glycol-trans-cyclooctene (TCO-PEG4-NHS) and N-hydroxysuccinimide ester-polyethylene glycol-tetraazine (Tz-PEG4-NHS) were purchased from Ruixi Biotechnology Co., Ltd. (Xi'an). 1-Ethyl-3-(3-dimethylaminopropyl)-carbodiimide hydrochloride (EDC), sodium N-hydroxysulfonated succinimide (Sulfo-NHS), phosphate-buffered saline (PBS), and 2-(N-morpholino)ethanesulfonate hydrate (MES) were purchased from Aladdin (Shanghai).

[0057] Bovine serum albumin (BSA) was purchased from Amresco (USA).

[0058] Chloramphenicol (CAP), neomycin, and clarithromycin were purchased from Sigma Aldrich (USA).

[0059] Chloramphenicol antibody (CAP-Ab) and BSA-chloramphenicol antigen (BSA-CAP-Ag) were purchased from Sangon Biotech Co., Ltd. (Shanghai).

[0060] The laboratory water was deionized using a water purification system (Millipore, USA). All chemicals were of analytical grade and required no further purification before use.

[0061] The preparation method of the relevant reagents used in this embodiment: PBS buffer (10 mM, pH=7.4): Take 8.00 g NaCl, 0.20 g KCl, 0.20 g KH2PO4 and 2.90 g Na2HPO4·12H2O and dilute to volume in a 1000 mL volumetric flask, then shake well.

[0062] MES buffer (0.1 M, pH=6.0): Dissolve 21.325 g of MES in deionized water and bring the volume to 1000 mL to obtain solution A; dissolve 4 g of NaOH in deionized water and bring the volume to 1000 mL to obtain solution B; mix 1000 mL of solution A and 400 mL of solution B and shake well.

[0063] PBST and MEST solutions: Add 0.5 mL of Tween-20 to 1000 mL of prepared PBS or MES buffer and shake well.

[0064] (1) Preparation of PS-CAP-Ab probe at room temperature Surface-functionalized carboxylic acid PS-6 μm microspheres (2 mg) were washed twice with MES buffer and then resuspended in MES. EDC (5 mg / mL, 30 mL) and Sulfo-NHS (5 mg / mL, 15 mL) (freshly prepared) were then added, and the mixture was incubated at room temperature for 15 minutes. After the reaction, the microspheres were washed with a centrifuge (5000 rpm) and resuspended in PBS. Subsequently, CAP-Ab (3.5 mg / mL, 100 μL) was added to the microspheres, and the coupling reaction was carried out at room temperature for 3 hours. After the coupling step, the non-specific binding sites on the microsphere surface were blocked with blocking buffer, and the coupled microspheres were washed three times with PBST. They were then stored below 4°C for later use.

[0065] (2) Preparation of MNP-BSA-CAP probe at room temperature 500 µg of surface-functionalized carboxylic acid MNP-150 nm particles were washed twice with MES buffer and then resuspended in MES. EDC (5 mg / mL, 30 mL) and Sulfo-NHS (5 mg / mL, 15 mL) (prepared fresh) were then added, and the mixture was incubated at room temperature for 15 minutes. After the reaction was complete, the particles were washed by magnetic separation and resuspended in PBS. BSA-CAP (50 µg) was then added, and the coupling reaction was carried out at room temperature for 3 hours. After the reaction was complete, the reaction mixture was purified by magnetic separation, and non-specific binding sites on the particle surface were blocked for 30 minutes with BSA (1%, 100 µL) blocking solution. After the blocking step, the particles were washed three times with PBST and stored at 4°C for future use.

[0066] (3) Chloramphenicol detection process Chloramphenicol standard solution (10 mg / mL, solvent: methanol) was serially diluted with PBS to prepare chloramphenicol solutions with concentrations ranging from 0 to 10 µg / mL. Pre-prepared chloramphenicol antibody-modified polystyrene microspheres (PSM-CAP-Ab, 2 mg / mL, 4 μL) and chloramphenicol complete antigen-modified magnetic particles (MNP-BSA-CAP, 500 µg / mL, 10 μL) were mixed with 150 μL of the diluted chloramphenicol standard solutions of different concentrations. The reaction was allowed to proceed via rotation at room temperature for 15 minutes to induce a competitive reaction. After the reaction, the supernatant was collected by magnetic separation and loaded into a deeply recessed glass slide sample chamber. The sample was then imaged using a holographic imaging system, and the resulting images were used for further data analysis.

[0067] The correlation between CAP concentration (high, medium, low) and PSM quantity indicates that the three-dimensional distribution of PSM in large-volume samples was successfully reconstructed. Figure 8 The concentration gradient response and linear fitting within the CAP concentration range (Figures 9a and 9b) show that the linear detection range of the 3D_EEA biosensor is 5 pg / mL to 100 ng / mL. The log-log linear regression equation between PSM quantity and chloramphenicol concentration is: log ( y ) = 0.207 log ( x ) + 2.143, and within this linear range, the coefficient of determination (R²) is as high as 0.988.

[0068] This invention further evaluated the specificity of the 3D_EEA biosensor in detecting CAP using clarithromycin and neomycin as interfering antibiotics. Figure 10The results showed no statistically significant difference between the control group and the interference group, while a significant difference existed between the control group and the chloramphenicol-containing group. These experimental results demonstrate that the 3D_EEA biosensor possesses excellent specificity and strong anti-interference capability.

[0069] Example 4: Experimental Protocol for Detecting Salmonella Using a 3D_EEA Biosensor This embodiment provides a holographic three-dimensional bioassay method based on depth tag encoding and attention mechanisms for detecting the concentration of Salmonella in a test sample. Salmonella is a common foodborne pathogen, and its detection is of great significance in food safety monitoring.

[0070] Required reagents and raw materials: Carboxyl-functionalized polystyrene microspheres (PSM-COOH, 6 μm), carboxylated magnetic nanoparticles (MNP-COOH, 150 nm), Salmonella polyclonal antibody (Salmonella-Ab), Salmonella surface antigen (Salmonella-Ag), 1-ethyl-3-(3-dimethylaminopropyl)-carbodiimide hydrochloride (EDC), N-hydroxysulfonated succinimide sodium salt (Sulfo-NHS), phosphate-buffered saline (PBS), 2-(N-morpholino)ethanesulfonate hydrate (MES), bovine serum albumin (BSA), and Salmonella standard strain (ATCC 14028).

[0071] The laboratory water was deionized using a Millipore water purification system. All chemicals were of analytical grade and required no further purification before use.

[0072] (1) Preparation of PSM-Salmonella-Ab probe at room temperature Carboxylic acid-functionalized PSM microspheres (2 mg) were washed twice with MES buffer and resuspended in MES. EDC (5 mg / mL, 30 μL) and Sulfo-NHS (5 mg / mL, 15 μL) (prepared fresh) were added, and the mixture was incubated at room temperature for 15 minutes. After the reaction, the microspheres were washed by centrifugation (5000 rpm) and resuspended in PBS. Salmonella antibody (Salmonella-Ab, 2 mg / mL, 100 μL) was added, and the coupling reaction was carried out at room temperature for 3 hours. The non-specific binding sites on the microsphere surface were blocked with BSA (1%, 100 μL) for 30 minutes. Finally, the microspheres were washed three times with PBST and stored at 4°C for later use.

[0073] (2) Preparation of MNP-Salmonella-Ag probe at room temperature Take 500 μg of surface carboxylic acid-functionalized MNP particles, wash twice with MES buffer, and resuspend in MES. Add EDC (5 mg / mL, 30 μL) and Sulfo-NHS (5 mg / mL, 15 μL) (prepared fresh for use), and incubate at room temperature for 15 minutes. After magnetic separation and washing, resuspend in PBS. Add Salmonella antigen (Salmonella-Ag, 50 μg), and couple the reaction at room temperature for 3 hours. After magnetic separation and purification, block with BSA (1%, 100 μL) for 30 minutes, wash three times with PBST, and store at 4°C for later use.

[0074] (3) Salmonella detection process Salmonella standard strains were serially diluted with PBS to prepare concentrations of 10-10. 6 CFU / mL bacterial suspensions were prepared. Pre-prepared PSM-Salmonella-Ab (2 mg / mL, 4 μL) and MNP-Salmonella-Ag (500 μg / mL, 10 μL) were mixed with 150 μL of Salmonella standard solutions of different concentrations. A competitive immunoreaction was performed by rotating the mixture at room temperature for 15 minutes. After the reaction, the supernatant was collected by magnetic separation and loaded onto a deeply grooved slide (depth 0.8 mm, diameter 15 mm). Holographic imaging was performed using a portable lensless holographic imaging microscope (light source wavelength 532 nm, CMOS pixel size 1.85 μm, resolution 4032×3036). The acquired two-dimensional holograms were reconstructed and particle counted using a 3D_EEA network. Finally, the bacterial concentration was calculated based on the linear relationship between PSM quantity and Salmonella concentration.

[0075] Example 5: Detection of chloramphenicol content in aquatic product samples.

[0076] The aquatic product (sea bass) samples were randomly purchased from the supermarket.

[0077] Pretreatment: Sample pretreatment was performed according to the national standard test method (GB 31658.2-2021). Specifically, the fish meat sample was pulverized using a grinder. Chloramphenicol solution of unknown concentration was randomly added to the pulverized fish meat, mixed thoroughly, and stored overnight at 4°C. Ultrasonic extraction was performed using a mixture of ethyl acetate and acetonitrile (1:1, v / v) for two cycles, each lasting 30 minutes. The resulting supernatant was evaporated under a nitrogen stream and resuspended in PBS solution for subsequent detection experiments.

[0078] The 3D_EEA biosensor was validated using real samples, with freshwater bass as the detection matrix. Due to the stringent quality control practices in aquaculture and the confirmation of antibiotic residues in the samples, six blind samples were artificially prepared by randomly adding different concentrations of CAP. Results ( Figure 11 The results showed that the number of positive samples detected by the 3D_EEA biosensor was consistent with the number of positive samples added to the blind samples. Based on the negative results of the parallel samples, sample number 8 was determined to be a false positive. Furthermore, comparison with enzyme-linked immunosorbent assay (ELISA) showed a high degree of consistency in detecting positive samples with high concentrations of chloramphenicol, and the 3D_EEA biosensor's limit of quantitation (LOQ) was lower than that of ELISA. These findings confirm the excellent performance of the 3D_EEA biosensor in practical sample analysis.

[0079] This invention develops a 3D_EEA biosensor that synergistically integrates the extended depth-of-field advantage of lensless holographic imaging with an attention-enhanced 3D depth reconstruction model through a large-volume counting strategy. By reconstructing traditional backpropagation or generative-based volume reconstruction algorithms into a depth classification-based framework, this model achieves higher computational efficiency and robust 3D spatial analysis of complex holographic signals. Furthermore, the deeply concave slide expands the effective detection volume and sampling depth. When combined with the 3D holographic reconstruction algorithm, this device increases the number of imaging signals detected while improving sensitivity and signal stability in large-volume biosensing. This approach not only addresses the key challenge of trace antibiotic detection but also provides a general framework applicable to various complex biosensing scenarios requiring precise spatial analysis. In future work, we aim to utilize this spatial framework for in-situ and high-throughput monitoring.

[0080] The above embodiments are merely preferred technical solutions of the present invention and should not be considered as limitations on the present invention. The embodiments and features described in these embodiments can be arbitrarily combined without conflict. The scope of protection of the present invention should be limited to the technical solutions described in the claims, including equivalent substitutions of the technical features described in the claims. That is, equivalent substitutions and improvements within this scope are also within the scope of protection of the present invention.

Claims

1. A holographic three-dimensional biometric method based on depth label encoding and attention mechanism, characterized in that: The method includes the following steps: Step 1: First, the target analyte is co-incubated with trans-cyclooctene-antibody conjugate and fully antigen-modified magnetic nanoparticles to perform an immunoassay. After the reaction, magnetic separation is performed, and the precipitate is collected. Subsequently, this precipitate is subjected to a bioorthogonal click cycloaddition reaction with tetrazine-functionalized polystyrene microspheres at room temperature. After the reaction, magnetic separation is performed again, and the supernatant obtained from this separation is finally collected and added to a custom-designed deep-grooved glass slide. Holographic imaging analysis is performed using a portable lensless holographic imaging microscope. The concentration of polystyrene microspheres in the supernatant is linearly correlated with the original concentration of the target analyte. Step 2: Using a random particle generation algorithm, a particle distribution with precise three-dimensional coordinate labels was created; then, the angular spectrum method combined with double Fourier transform was used to simulate the forward digital holographic propagation process, and Gaussian noise of different intensities was added to simulate real-world holographic data, resulting in a dataset containing three-dimensional coordinates and corresponding two-dimensional holograms. Step 3: An end-to-end network architecture based on the attention mechanism, 3D_EEA network, is used to complete the 3D reconstruction of the 2D holographic image. By using a 3D depth dataset and combining it with a deep learning model, coordinate localization and depth regression are performed on particles at different depths to complete the localization and reconstruction of particles in 3D space. Step 4: Utilizing the linear relationship between the number of polystyrene microspheres and the concentration of the target analyte, the polystyrene microspheres are located and quantitatively analyzed through holographic three-dimensional reconstruction. Finally, the number of polystyrene microspheres is calculated to obtain the concentration information of the target analyte.

2. The holographic three-dimensional biometric method based on depth label encoding and attention mechanism according to claim 1, characterized in that: Step 2 specifically includes the following: 2.1 A synthetic particle field is generated in three-dimensional space using a random particle generation algorithm; the spatial coordinates (X, Y, Z) of the particle center are generated according to the Poisson distribution, while allowing 30% particle overlap along the Z-axis to simulate the random and independent distribution of particles in a large volume space under real conditions. The coordinate information of each particle includes the bounding box coordinates (x, y, w, h) and the normalized depth label z ∈ [0,1], where z is discrete into K=100 categories, which constitute the depth classification label; 2.2 The three-dimensional particle distribution is backpropagated to the XY plane using the angular spectrum iterative method to synthesize the corresponding two-dimensional hologram. Specifically, this includes: initializing the light field, adding random particles, performing Fourier transform and phase modulation using the angular spectrum transfer function, and recording the superimposed hologram. To simulate a real imaging environment, Gaussian noise with different signal-to-noise ratios was added to the synthetic holograms; finally, a dataset containing 3D coordinate labels and corresponding 2D holograms was constructed to train an end-to-end 3D reconstruction model. In step 2, the dataset of two-dimensional holograms includes an XML file with particle localization and depth label information, as well as the corresponding digital holograms; a total of 25,000 samples are generated in the dataset and divided into training set, validation set and test set in a ratio of 7:2:1; to simulate real-world scenarios and achieve data augmentation, Gaussian noise with different signal-to-noise ratios is added to 5,000 images in the training set; in addition, 200 experimentally acquired holograms are included in the test set to evaluate the predictive performance of the 3D_EEA model on simulated data and real data.

3. The holographic three-dimensional biometric method based on depth label encoding and attention mechanism according to claim 2, characterized in that: In step 2.2, the angular spectrum iterative algorithm propagates the three-dimensional particle distribution into the two-dimensional holographic plane. The light source wavelength λ is set to 633 nm, the imaging sensor pixel size (Δx, Δy) is set to 1.85 μm, the refractive index n of the propagation medium is 1.33, the propagation distance (z) is 0.5 mm-1.5 mm, the calculation grid is set to 512×512 pixels, and the angular spectrum transfer function is defined as: H(fx, fy) = exp( j*2π * (z / n) * sqrt( (1 / λ)² - fx² - fy² ) ); the simulated particle diameter is the same as the actual particle diameter, which is 6 μm.

4. The holographic three-dimensional biometric method based on depth label encoding and attention mechanism according to claim 2, characterized in that: In step 2.2, Gaussian noise with different signal-to-noise ratios is added. Specifically, the signal-to-noise ratios are 10 dB, 15 dB, and 20 dB.

5. The holographic three-dimensional biometric method based on depth label encoding and attention mechanism according to claim 1, characterized in that: In step S3, the 3D_EEA network adopts an encoder-decoder architecture and embeds an attention mechanism. Its specific structure and data flow are as follows: The feature extraction encoder consists of four lightweight residual blocks connected in series, which serve as the backbone of the network. It is responsible for extracting multi-level features from the input two-dimensional hologram and outputting a high-dimensional feature tensor. The multi-scale feature fusion module consists of an upsampling path UF module and a downsampling path DF module. This module receives multi-level outputs from the encoder and forms a feature pyramid structure through cross-layer cascading. Finally, it outputs an enhanced feature that integrates multi-scale information, which is used to fuse the deep semantic features and shallow spatial detail features output by the encoder to improve the perception of particles at different scales. The convolutional attention module performs adaptive feature redefinition of the fused feature map F2 in terms of channel and spatial dimensions through the channel attention sub-module and the spatial attention sub-module, outputs the redefinition features, enhances the network's attention to particle regions, and suppresses the interference of complex background noise. The output head consists of a deep regression branch and a counting branch. Based on the redefinition features, the network finally outputs the three-dimensional coordinates (x, y, z), confidence score, and total number of particles in the image for each particle. The network training adopts a supervised learning approach, using the dataset constructed in step 2. The loss function includes target loss, bounding box regression loss, and depth regression loss. This network architecture design enables holographic three-dimensional spatial localization and quantitative analysis of particles.

6. The holographic three-dimensional biometric method based on depth label encoding and attention mechanism according to claim 5, characterized in that: In step S3, the channel attention submodule first generates two one-dimensional 1D channel descriptors through global average pooling and global max pooling. Then, these descriptors are processed by a shared multilayer perceptron (MLP) to generate channel attention weights. The MLP contains two convolutional layers with a kernel size of 1. The specific process is as follows: the first layer reduces the channel dimension by a reduction factor r=16, followed by a ReLU activation operation; the second layer restores the original channel dimension; and the final output weights are normalized using the sigmoid function.

7. The holographic three-dimensional biometric method based on depth label encoding and attention mechanism according to claim 5, characterized in that: In step S3, the spatial attention submodule takes the channel-optimized features as input and first performs average pooling and max pooling operations along the channel dimension to generate two two-dimensional 2D spatial feature maps. These feature maps are then concatenated and processed through a standard convolutional layer to generate a spatial attention map. The specific parameters of this convolutional layer are as follows: number of input channels = 2; number of output channels = 1; kernel size = 7×7; stride = 1; padding = 3. The resulting spatial weights are also normalized using the sigmoid function.

8. The holographic three-dimensional biometric method based on depth label encoding and attention mechanism according to claim 4, characterized in that: In step S3, the network training parameters of the end-to-end 3D reconstruction model include: 200 training iterations, an initial learning rate of 0.01, a batch size of 16, using the Adam optimizer, a momentum of 0.937, and a weight decay of 0.0005. In step S3, the loss function is specifically as follows: ； in, For bounding box regression loss, For deep classification, cross-entropy loss, The target confidence binary cross-entropy loss is used. The value is 0.

25. The value is 1.

0. The value is set to 1.0, and the weight parameter is determined through empirical parameter tuning.

9. The holographic three-dimensional biometric method based on depth label encoding and attention mechanism according to claim 1, characterized in that: The portable lensless holographic imaging microscope in step S1 includes a coherent light source, an optical slit, a CMOS image sensor, and a 3D-printed dark-field shell. The wavelength of the coherent light source is 532 nm, the pixel size of the CMOS image sensor is 1.85 μm, and the resolution is 4032×3036. The depth of the recessed slide is 0.8 mm, the diameter is 15 mm, and the thickness is 1 mm. PSM consists of carboxylated polystyrene microspheres with a diameter of 6 μm, and the surface is modified with Tz click molecules for click reactions. MNPs are carboxylated magnetic nanoparticles with a diameter of 150 nm, and their surfaces are modified with the complete antigen of the target analyte. The target analyte is a small molecule antibiotic, including chloramphenicol, neomycin, clarithromycin or other antibiotics and pathogenic bacteria. The holographic immunoassay method has a linear detection range of 5 pg / mL to 100 ng / mL for chloramphenicol, with a coefficient of determination R² of 0.

988.

10. The holographic three-dimensional biometric method based on depth label encoding and attention mechanism according to claim 1, characterized in that: The time required for step S1, the holographic immunoassay method, is 40-50 minutes. The reaction conditions for the competitive immunoassay are: 20-40 ℃, and the reaction time is 10 min.

Images